AI Foundation & Prompting
AI के हो, कसरी सिक्छ, कहाँ उपयोगी हुन्छ, कहाँ सावधानी चाहिन्छ—र राम्रो prompt कसरी लेख्ने भन्ने विस्तृत beginner-friendly note.
AI, ML, LLM, token, context र agent बुझ्ने।
आफ्नो role का लागि practical use case छान्ने।
Clear context र format सहित instruction लेख्ने।
Facts, privacy, bias र uncertainty जाँच्ने।
AI Foundation
पहिले foundation स्पष्ट बनाऔँ: intelligence के हो, AI के हो, machine कसरी सिक्छ, र modern AI terminology को अर्थ के हो।
1.1 मानव बुद्धि: Think → Learn → Decide
मानव intelligence को मूल चक्र भनेको information बुझ्ने, experience बाट सिक्ने, र context तथा values प्रयोग गरेर निर्णय गर्ने क्षमता हो।
सोच्ने · THINK
Information लाई केवल देख्नु होइन—त्यसको अर्थ, सम्बन्ध र implication बुझ्नु।
सिक्ने · LEARN
Experience, feedback र mistake बाट आफ्नो approach सुधार्नु।
निर्णय · DECIDE
Goal, context, values, risk र consequence हेरेर action छान्नु।
1.2 AI भनेको के हो?
सरल परिभाषा: Artificial Intelligence भनेको data वा examples बाट pattern सिकेर prediction, recommendation, classification वा नयाँ content तयार गर्ने machine-based system हो।
महत्त्वपूर्ण: AI “मानिसजस्तै जान्ने चेतनशील प्राणी” होइन। यसको output learned patterns, supplied context र system rules बाट आउँछ।
1.3 AI vs Human Intelligence
AI को strength
- धेरै information छिटो process गर्ने
- Large dataset मा pattern खोज्ने
- Repeatable task consistent बनाउने
- Draft, summarize, classify र transform गर्ने
- तर lived experience वा personal accountability हुँदैन
Human को strength
- Empathy, values र ethical judgment
- Local culture, nuance र unstated context
- Responsibility र consequence बुझ्ने
- Purposeful creativity र original intent
- Ambiguous situation मा accountable decision
Best work = AI speed × Human judgment. AI लाई replacement होइन, capability amplifier को रूपमा बुझ्नु उपयोगी हुन्छ।
1.4 दैनिक जीवनमा AI कहाँ छ?
Face Unlock
Camera image बाट facial pattern match गरेर device unlock गर्छ।
Google Maps
Traffic, distance र historical route data बाट travel time र route सुझाव दिन्छ।
YouTube Feed
Watch history र engagement pattern अनुसार recommendation rank गर्छ।
Search Results
Query intent र web signals प्रयोग गरेर relevant result क्रम मिलाउँछ।
Fraud Detection
Unusual transaction वा behavior pattern detect गरेर alert दिन्छ।
Social Feed
Content, people र ads लाई predicted interest अनुसार rank गर्छ।
1.5 AI Family Tree
यी शब्द एउटै कुरा होइनन्। तिनीहरू व्यापक field बाट specific technology तर्फ जाने hierarchy हुन्।
1.6 Essential AI Glossary
Filter गरेर term हेर्नुहोस्। प्रत्येक term सँग beginner-friendly explanation र Nepali-context example दिइएको छ।
मानव intelligence चाहिने जस्तो देखिने task गर्ने broad field.
Example: banking fraud alertExplicit rule मात्र होइन, data बाट pattern सिक्ने AI method.
Example: spam email filterधेरै processing layers प्रयोग गरेर complex pattern सिक्ने ML approach.
Example: speech recognitionConnected mathematical nodes/weights बाट input transform गर्ने model structure.
Example: photo classificationHuman language analyze, understand वा generate गर्ने technology.
Example: Nepali text summaryImage वा video बाट object, face, text वा event बुझ्ने system.
Example: face unlockLearned pattern प्रयोग गरेर नयाँ text, image, audio, code वा video बनाउने AI.
Example: email first draftLarge amount of text मा train भएको language prediction/generation model.
Example: conversational assistantModel ले process गर्ने text को सानो unit—word, subword वा character को भाग.
Used for: context and usage limitsAI लाई दिइने instruction, question, context वा task brief.
Example: “Write a client follow-up email…”Model ले एक interaction मा ध्यान दिन सक्ने total input/output information को सीमा.
Example: long PDF analysis limitAI ले fluent र confident तर गलत fact, source वा detail बनाउनु.
Example: non-existent citationTraining data, design वा use context का कारण skewed वा unfair result आउनु.
Example: unfair candidate rankingAnswer दिनुअघि trusted documents वा knowledge source retrieve गर्ने approach.
Example: company policy Q&AGoal पूरा गर्न planning, tools र multiple steps प्रयोग गर्ने AI-based system.
Example: research → summarize → draftDefined trigger र workflow अनुसार repeated task स्वतः चलाउनु.
Example: form → CRM → emailWhy AI Matters in 2026
Capability, adoption र access बढेका छन्। तर AI access हुनु र organizational value निकाल्नु फरक कुरा हो।
2.1 Global AI Landscape
Access ≠ integration. Sustainable value आउन workflow redesign, governance, training र human accountability चाहिन्छ।
Sources: Stanford AI Index 2026 and McKinsey State of AI 2025. These global figures should not be treated as Nepal-specific statistics.
2.2 AI coding का लागि मात्र होइन
Writing
Email, report, proposal, translation, editing.
Learning
Explanation, questions, feedback, study plans.
Office Work
Summary, meeting notes, SOP and structure.
Marketing
Idea, campaign angle, copy and repurposing.
Design
Brief, visual direction, presentation draft.
Business
Research, offer, operations and support.
Customer Support
FAQ, routing, first reply and escalation.
Decision Support
Compare options—while humans decide.
2.3 Nepal Context
नेपालमा AI को value “सबैभन्दा advanced model” बाट मात्र आउँदैन। Local language, practical workflow, affordable access र trusted sources महत्वपूर्ण हुन्छन्।
Teachers & Schools
Lesson support, learning materials, simplified explanation—student data सुरक्षा सहित।
Local Government
Official notice लाई simple language मा explain वा FAQ draft—policy invent नगरी।
Banking & Finance
Fraud awareness, support routing, document summary—strict privacy control सहित।
Digital Services
Nagarik App/SSF-style service instructions लाई accessible language मा explain गर्ने।
Creators
Nepali content idea, script, caption, repurposing र audience adaptation.
SMEs & Entrepreneurs
Customer research, offer writing, marketing operations र repeat support.
2.4 AI as Infrastructure
AI एउटा isolated app मात्र होइन। उपयोगी system मा user goal, workflow, AI capability, context र data जोडिन्छन्।
AI optional skill मात्र होइन—digital workflow सँग जोडिँदै गएको infrastructure layer हो।
Practical AI Use Cases
Use case छान्दा “कुन tool?” भन्दा पहिले “कुन repeated pain point वा outcome?” सोध्नुहोस्।
STUDENT
Use AI for
Concept explanation, practice questions, study plan, feedback, comparison and revision support.
Healthy learning loop
Ask → Try yourself → Check → Explain in your own words.
TEACHER
Use AI for
Lesson outline, differentiated activities, examples, question bank, rubric draft and parent communication.
Human role
Adapt by grade, culture and learner needs. Teacher makes final instructional and assessment decisions.
OFFICE
Use AI for
Email drafts, report summary, meeting minutes, action extraction, SOP outline and presentation structure.
Workflow
Input → AI draft → fact/policy/tone review → final send or filing.
CREATOR
Use AI for
Idea bank, hooks, script variations, captions, content calendar and long-to-short repurposing.
Creative shift
Blank page → multiple options → human voice, taste and final story.
ENTREPRENEUR
Use AI for
Customer research, offer draft, competitor comparison, campaign planning, support FAQ and review analysis.
Best starting point
One repeated customer or operational pain point—not twenty disconnected AI tools.
FREELANCER
Use AI for
Proposal draft, discovery questions, research, project checklist, status update and quality review.
Advantage
Speed + quality + communication, supported by reusable workflows and custom judgment.
3.1 Public Information & Customer Support
✓ Good public-information use
- Official notice translate वा simplify गर्ने
- Approved document बाट FAQ draft गर्ने
- Process लाई step-by-step explain गर्ने
- Official page वा notice को link राख्ने
✕ Never invent
- Eligibility वा legal requirement अनुमान गर्ने
- Deadline वा fee guess गर्ने
- Unofficial policy लाई official जस्तो प्रस्तुत गर्ने
- High-impact case मा human escalation हटाउने
3.2 Research, Learning & Presentation Tools
Research workflow
Collect PDFs/links → Ask focused questions → Trace citations → Synthesize in your own reasoning. Notebook-style tools and answer engines are valuable when sources remain visible.
Presentation workflow
Brief → structure → visual draft → edit → present. Gamma, Canva and Napkin-style tools can speed up first drafts, but story, accuracy and final design judgment remain human work.
AI Safety & Responsible Use
Fluent output truth को guarantee होइन। Safety लाई one-time warning होइन, repeatable workflow बनाउनुहोस्।
4.1 Four Common AI Limitations
⚠ Hallucination
AI ले fake fact, quote, source वा detail confidently बनाउन सक्छ।
◷ Outdated Data
Model knowledge current नहुन सक्छ; live information verification चाहिन्छ।
⚖ Bias
Training data वा design choices बाट skewed, incomplete वा unfair output आउन सक्छ।
▣ Privacy Risk
Prompt मा राखिएको information तपाईंको direct control बाहिर जान सक्छ।
Fluent language is not proof. Output राम्रो सुनिनु र output सत्य हुनु दुई अलग कुरा हुन्।
4.2 Sensitive Data Boundary
✕ Never put in public AI
- Citizenship, passport वा national ID number
- Bank account, card details वा transaction credentials
- Password, PIN, OTP वा recovery code
- Private client file वा unpublished contract
- Medical, HR वा student records
- Confidential source code वा company strategy
✓ Safer alternatives
- Name र identifier हटाएर anonymize गर्ने
- Real data को सट्टा sample data प्रयोग गर्ने
- Approved enterprise account वा tool प्रयोग गर्ने
- Company/school policy check गर्ने
- Permission लिने र minimum data मात्र दिने
- Highly sensitive summary locally तयार गर्ने
4.3 Human-in-the-Loop
4.4 NIST-Inspired Safety Loop
GOVERN
Rules, roles, responsibility र accountability define गर्ने।
responsible use
MANAGE
Risk reduce, monitor, respond र continually improve गर्ने।
MAP
Use case, affected people, context र possible harm बुझ्ने।
MEASURE
Accuracy, bias, privacy, robustness र failure test गर्ने।
Based on NIST AI Risk Management Framework 1.0; simplified for beginner training.
4.5 Deepfake & Scam Awareness
1. Pause
Urgency, fear वा “अहिले नै पैसा/OTP पठाउनु” भन्ने pressure मा तुरुन्त action नगर्नुहोस्।
2. Inspect
Voice quality, lip sync, strange wording, account history, number र link domain जाँच्नुहोस्।
3. Verify
Known phone number, official website वा trusted person मार्फत independently confirm गर्नुहोस्।
Prompting: AI बाट Useful Output निकाल्ने Skill
Prompt भनेको magic sentence होइन। यो goal, context, examples, rules र quality checks भएको work brief हो।
5.1 Prompt भनेको के हो?
Prompt = AI लाई दिइने कामको brief. तपाईंले के चाहनुहुन्छ, कसका लागि, कुन information प्रयोग गरेर, कस्तो format मा, कुन सीमाभित्र भन्ने instruction हो।
You = Manager
- Goal स्पष्ट गर्ने
- Relevant context दिने
- Success criteria define गर्ने
- Final quality check गर्ने
AI = Smart Intern
- Fast first draft बनाउने
- Examples र pattern follow गर्ने
- Missing context मा गलत अनुमान गर्न सक्ने
- Clarification र review चाहिने
5.2 Bad → Better → Best Prompt
Aligned with current Google prompt design and Anthropic prompt engineering guidance.
5.3 Complete Prompt Formula
You are a [ROLE]. Your task is to [TASK]. Context: - Audience: [WHO] - Goal: [WHY] - Source information: [INPUT] Return the result as [FORMAT]. Follow this example or pattern: [EXAMPLE]. Constraints: - [LENGTH / TONE / RULE] - Do not invent missing information. - Mention uncertainty clearly. - Ask questions first if critical context is missing.
5.4 Simple Framework: Persona + Task + Context + Format
छोटो prompt को लागि चार-part minimum structure प्रयोग गर्न सकिन्छ:
Specific input → useful output. धेरै शब्द लेख्नु नै राम्रो prompting होइन; relevant detail स्पष्ट लेख्नु राम्रो prompting हो।
5.5 Copyable Live-Demo Prompts
Demo 1 · Business Email
Recipient, tone, context, constraint र CTA स्पष्ट गर्नुहोस्।
You are a professional business writer. Write a warm follow-up email to Hari after our digital marketing proposal meeting. Mention the Rs. 25,000 monthly package. Maximum 130 words. Include a subject line. End with one clear next step.
Demo 2 · Facebook Post
Exact style imitation भन्दा voice attributes define गर्नु सुरक्षित र repeatable हुन्छ।
Create a Nepali-English Facebook post for Rabin Paudel. Voice: practical, warm, beginner-friendly, short sentences, one local example. Topic: AI ले job replace होइन, work upgrade गर्न सक्छ. Audience: Nepali students and office professionals. Length: 120–150 words. End with a thoughtful question.
Demo 3 · Teacher Lesson Plan
Grade, objective, time, activities र assessment ले plan usable बनाउँछ।
You are a Grade 8 science teacher in Nepal. Create a 40-minute lesson plan on photosynthesis. Include: 1. Learning objective 2. Required materials 3. Three learning activities 4. Differentiation for mixed ability 5. A 5-minute exit ticket Return the result as a table.
Demo 4 · Content Creator Script
Hook, audience, duration, structure र CTA specify गर्नुहोस्।
Write a 60-second Nepali-English video script for small business owners. Topic: 3 safe ways to use AI for marketing. Start with a surprising 3-second hook. Use three numbered points and one simple Nepali example. End with: “पहिले verify, अनि publish.”
Demo 5 · Office Report Summary
Source, decision need, output format र no-invention rule जोड्नुहोस्।
Summarize the attached weekly operations report for the manager. Return: - 3 wins - 3 risks - 3 decisions needed - Owners and deadlines, only if stated Do not invent missing data. Separate facts from assumptions. Mark uncertainty clearly.
5.6 Advanced Prompting Techniques
Few-Shot Prompting · Examples देखाएर pattern सिकाउने
One or more input-output examples दिएर AI लाई desired style, classification rule वा format देखाइन्छ। यो tone matching, labeling, custom formatting र consistent output मा useful हुन्छ।
Classify each post: Example 1: Input: “आज पानी पर्यो” Output: “Weather update” Example 2: Input: “नयाँ offer सुरु” Output: “Promotion” Now classify: Input: “Workshop भोलि बिहान १० बजे”
Persona Prompting · Perspective define गर्ने
“Act as a teacher/editor/consultant/customer” भनेर output को lens define गर्न सकिन्छ। Persona facts को replacement होइन; role जस्तोसुकै भए पनि verification उस्तै आवश्यक हुन्छ।
Multi-Step Prompting · Complex task लाई stages मा बाँड्ने
Goal define → outline → draft → critique → revise. प्रत्येक stage review गर्न सकिने भएकाले one giant prompt भन्दा control र quality राम्रो हुन सक्छ।
Prompt Chaining · एक output अर्को चरणको input
Research findings → outline → draft → critique → final revision. Chain को प्रत्येक handoff मा required format define गर्दा workflow predictable हुन्छ।
5.7 Verification Prompting
Before finalizing: 1. Check every factual claim. 2. Cite reliable sources with publication dates. 3. Separate facts from assumptions. 4. Mention what you are uncertain about. 5. Do not invent citations, links, quotes or statistics. 6. If current information is required, say whether you verified it live.
5.8 JSON Prompting for Automation
Structured output लाई forms, CRM, database, Zapier/Make-style automation वा API workflow मा reuse गर्न सजिलो हुन्छ।
{
"customer_name": "string",
"intent": "sales | support | complaint",
"priority": "low | medium | high",
"summary": "string",
"next_action": "string"
}
For complex production systems, prefer a provider's structured-output/schema feature when available; prompt-only JSON can still require validation.
5.9 Context Engineering
Prompt एउटा instruction हो। Context engineering भनेको AI ले राम्रो काम गर्न आवश्यक पूरा environment तयार गर्नु हो।
Modern AI work मा “perfect prompt” भन्दा “correct context + clear workflow + verification” धेरै महत्वपूर्ण हुन्छ।
Practice, Assignment & Recap
Learning personal तब हुन्छ जब तपाईं आफ्नो वास्तविक कामका लागि prompts बनाउनुहुन्छ।
6.1 Participant Exercise · 12 Minutes
Repeat Task
बारम्बार गर्ने एउटा काम छान्नुहोस्—email, summary, report, FAQ आदि।
Create Task
Script, post, lesson, proposal वा presentation outline जस्तो creation task.
Decision Support
Options compare, risks identify वा plan structure गर्ने task.
6.2 Day 1 Assignment
1. Personal Introduction
80–120 words. Audience र tone स्पष्ट गर्नुहोस्।
2. One Facebook Post
Topic, target audience, voice attributes र CTA राख्नुहोस्।
3. One Office Email
Recipient, context, action needed र maximum length define गर्नुहोस्।
4. One Learning Plan
7 days, daily action, milestones र review method सहित।
ChatGPT वा Gemini प्रयोग गर्नुहोस्। Final output मात्र होइन—तपाईंले प्रयोग गरेको prompt पनि save गर्नुहोस्।
6.3 Day 1 Recap
6.4 Day 2 Preview
Day 2 मा prompt skill लाई practical tool workflow मा जोडिन्छ: tool landscape, research workflows, content workflows, office automation र hands-on challenge.
Credible Sources
Statistics, safety structure र prompting guidance का लागि official or primary sources प्रयोग गरिएको छ।